高空间分辨率遥感影像分割尺度参数自动选择研究

Study on the Automatic Selection of Segmentation Scale Parameters for High Spatial Resolution Remote Sensing Images

面向对象解译技术在高分辨率遥感影像信息提取中得到广泛应用,但影像分割的基础问题仍严重制约其自动化水平,尤其是分割参数选择。因此,本文以广泛使用的分型网络演化分割算法为例,开展尺度参数选择研究。借鉴对遥感影像分辨率敏感的局部方差指标,引入边长和面积权重,构造加权局部方差(WLV)指标,对多个分割结果进行评价,进而实现最佳尺度参数选择。在珠江区域2.5 m的SPOT 5融合影像上进行实验,通过计算最佳分割结果与人工分割结果的相似度对WLV进行定量验证。此外,还对WLV在分割对象最小为一个像元、最大为整景影像的全范围尺度参数的变化规律进行了实验,结果表明:在WLV随尺度参数的变化曲线中,不同极大值点的分割结果反映了实验区不同景观层级上的斑块,其中第1个极大值点对应的分割结果能够较好地反映影像的最小可识别单元。

Geographic Object- Based Image Analysis（GEOBIA） is widely used in high spatial resolution remote sensing image interpretation. However, the fundamental component of image segmentation severely obstructs its development, especially on the selection of segmentation parameters. To overcome these issues, we choose the widely used Fractal Network Evolution Algorithm as an example, which is provided by e Cognition, and focus on looking for an approach for the scale parameter selection. Inspired by the similarity between the merging segmentation and the degrading image resolution that the unit size within both would increase,we introduce the weights of border length and the weights of object area into the metric of Local Variance proposed by Woodcock and Strahler（1987）, and propose a new segmentation evaluation metric： Weighted Local Variance（WLV）. Through comparing WLV with a supervised metric on a series of segmentations with limited increasing scale parameters, we found that the best segmentation result chosen by the first local maximum point of the scale-WLV curve is similar to the manual segmentation result.Then we validate WLV on two more images and expand the limited scale space to the full range, so that the segments can change from one pixel to the whole image. Results show that the segmentations chosen by WLV local maximum points could reflect the different levels in the hierarchical landscape, and the segmentation of the first levels is capable of expressing the finest homogeneous patches.